An Empirical Comparison of Instance Attribution Methods for NLP

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 967–975 June 6–11, 2021. ©2021 Association for Computational Linguistics 967 An Empirical Comparison of Instance Attribution Methods for NLP Pouya Pezeshkpour∗ University of California, Irvine pezeshkp@uci.edu Sarthak Jain∗ Northeastern University jain.sar@northeastern.edu Byron C. Wallace Northeastern University b.wallace@northeastern.edu Sameer Singh University of California, Irvine sameer@uci.edu Abstract Widespread adoption of deep models has motivated a pressing need for approaches to interpret network outputs and to facilitate model debugging. Instance attribution meth- ods constitute one means of accomplishing these goals by retrieving training instances that (may have) led to a particular prediction. Influence functions (IF; Koh and Liang 2017) provide machinery for doing this by quan- tifying the effect that perturbing individual train instances would have on a specific test prediction. However, even approximating the IF is computationally expensive, to the degree that may be prohibitive in many cases. Might simpler approaches (e.g., retrieving train examples most similar to a given test point) perform comparably? In this work, we evaluate the degree to which different potential instance attribution agree with respect to the importance of training samples. We find that simple retrieval methods yield training instances that differ from those identified via gradient-based methods (such as IFs), but that nonetheless exhibit desirable characteristics similar to more complex attribution methods. Code for all methods and experiments in this paper is available at: https://github.com/successar/ instance_attributions_NLP. 1 Introduction Interpretability methods are intended to help users understand model predictions (Ribeiro et al., 2016; Lundberg and Lee, 2017; Sundararajan et al., 2017; Gilpin et al., 2018). In machine learning broadly and NLP specifically, such methods have focused on feature-based explanations that highlight parts of inputs ‘responsible for’ the specific prediction. Feature attribution, however, does not communi- cate a key basis for model outputs: training data. Recent work has therefore considered methods for ∗Equal contribution A hilarious romantic comedy ˆy = 1 <latexit sha1_base64="0mhe76YrMN0t/3b7j+R/G9n/sU=">AB8HicbVBNS8NAEJ34WetX1aOXxSJ4KkV9CIUvXisYD+kDWz3bRLN5uwOxFK6K/w4kERr/4c b/4bt20O2vpg4PHeDPzgkQKg67aysrq1vbBa2its7u3v7pYPDpolTzXiDxTLW7YAaLoXiDRQoeTvRnEaB5K1gdDv1W09cGxGrBxwn3I/oQIlQMIpWeuwOKWbjybXK5XdijsDWSZeTsqQo94rfX7MUsjrpBJakzHcxP0M6pRMknxW5qeELZiA54x1JFI278bHbwhJxapU/CWNtSGbq74mMRsaMo8B2RhSHZtGbiv95nRTDKz8TKkmRKzZ fFKaSYEym35O+0JyhHFtCmRb2VsKGVFOGNqOiDcFbfHmZNKsV7xSvb8o127yOApwDCdwBh5cQg3uoA4NYBDBM7zCm6OdF+fd+Zi3rj5zBH8gfP5A6oJkFA=</latexit> input prediction Attribution methods e.g.: IF, GD, NN (dot) IF, GD: … interested in nothing more than sucking you in … Important positive examples Important negative examples NN (dot): if you can get past the taboo subject … REP: … a canny crowd pleaser … IF, GD, REP: eerily accurate depiction of depression. NN (dot): insufferably naive. Figure 1: Attribution methods score train examples in terms of their importance to a particular prediction. In this work, we compare several such methods, e.g., In- fluence Functions (IF) and its variants (GD), Represen- ter Points (REP) and similarity measures (NN). surfacing training examples that were influential for a specific prediction (Koh and Liang, 2017; Yeh et al., 2018; Pezeshkpour et al., 2019; Charpiat et al., 2019; Barshan et al., 2020; Han et al., 2020). While such instance-attribution methods provide an appealing mechanism to identify sources that led to specific predictions (which may reveal po- tentially problematic training examples), they have not yet been widely adopted, at least in part be- cause even approximating influence functions …

Content truncated to 500 words